Analyzing Plan Diagrams of Data Query Optimizers

Size: px

Start display at page:

Download "Analyzing Plan Diagrams of Data Query Optimizers"

Erik O’Brien’
8 years ago
Views:

1 Analyzing Plan Diagrams of Data Query Optimizers INSTITUT FÜR PROGRAMMSTRUKTUREN UND DATENORGANISATION (IPD), FAKULTY OF INFORMATICS 1 KIT University of the State of Baden-Wuerttemberg and National Research Center of the Helmholtz Association Name of Institute, Faculty, Department

2 Paper 2 Analyzing Plan Diagrams of Database Query Optimizers written in 2005 Naveen Reddy, Jayant R. Haritsa Database Systems Lab, SERC/CSA Indian Institute of Science, Bangalore , INDIA

3 Relational Database Management System Where is a Query Optimizer located? Relational Database Management System (RDBMS) Client SQL Query Optimizer Parser Result... Database 1 Database 2 Database 3 Database System 3

4 Query Optimizer How does a Query Optimizer work? Parsed Query (from Parser) subquery unnesting, query rewrite with materialized views,... Query Transformer Transformed Query Estimator Dictionary Query and estimates Plan Generator Query Plan (to Row Source Generator) 4

5 Query Optimizer 5 Identifies the most efficent strategy to execute SQL queries Returns query plan Part of the RDBMS Based on estimated number of rows of relevant relations Relevant join conditions are given in the query

6 Query Plan A strategy to execute SQL queries Query Plan does not affect the result Example: (id is FK of Aid) select o,p from A inner join B on (A.id=B.Aid) where B.p=0; Plan P1: AB A=3, B=6/3=2 => cost(ab)=6 id o X Y Z Plan P2: BA 6 A B id Aid p B=3 (using where), A=1 (primary key) => cost(ba)=3

7 How to analyze a query optimizer? 1. Send SQL statement EXPLAIN SELECT FROM WHERE 2. Optimizer returns query plan 3. This can be repeated by varying the where condition(s) systematically until the selectivity space is covered 4. Corresponding plan and cost for each set of conditions can be visualized Customers 4711 Selectivity Space 13 7 Orders

8 Query plan costs A id o X Y Z B id Aid p test=# explain select o,p from A inner join B on (A.id=B.Aid) where B.p=0; QUERY PLAN Hash Join (cost= rows=3 width=6) Hash Cond: (b.aid = a.id) -> Seq Scan on b (cost= rows=3 width=8) Filter: (p = 0) -> Hash (cost= rows=3 width=6) -> Seq Scan on a (cost= rows=3 width=6) 8

9 TPC-H Decision support benchmark database schema, data queries for testing (Q1 through Q22) We use same data set as the authors of the paper: relation part.tbl nation.tbl customer.tbl region.tbl partsupp.tbl supplier.tbl orders.tbl lineitem.tbl 9 #rows

10 Picasso 10 Picasso Sends explain queries against the database Stores and visualizes query plan and costs Picasso Server Connects to Database (SQL-Server, Oracle, Sybase, postgresql,...) Multiuser Picasso Client Connects to PICASSO Server GUI for PICASSO Server

11 Run test with Picasso 11

12 Run test with Picasso 12 Picasso varies WHERE conditions according to plot resolution: and o_totalprice :varies and c_acctbal :varies 2 varying conditions 2-dimensional selectivity space (2D) Picasso runs EXPLAIN-Statements to determine the query plan chosen by the optimizer. 2D and Resolution= ² = 10,000 EXPLAIN-Statements

13 Diagram creation process Create EXPLAIN SELECT... SQL-Statement with condition(s) SQL Optimizer RDBMS Add new point to the Plan Diagram Query Plan Output Diagram 13 explain select name, price from customer, order where [ ] and o_totalprice <= and c_acctbal <= 5809; result: P4

14 Plan Diagram and Cost Diagram Plan Coverage of selectivity space in % Plan Diagram: 2D-visualization of execution plans 14 Cost Diagram: 3D-visualization of estimated costs. Estimation is done by optimizer.

Cost domination Principle Monotonic cost behaviour: Increasing selectivity space increases estimated costs Non-monotonic cost behaviour: Increasing

15 Cost domination Principle Monotonic cost behaviour: Increasing selectivity space increases estimated costs Non-monotonic cost behaviour: Increasing selectivity space does not always increase estimated costs Cost domination principle: When increasing selectivity space, estimated costs must increase monotonically. 15

Estimated costs vs. execution costs SQL Statement for each point in diagram Measurement Creation time for a 10x10 diagram Monotonicity Estimated costs Execution costs explain select... select... #rows,.

16 Estimated costs vs. execution costs SQL Statement for each point in diagram Measurement Creation time for a 10x10 diagram Monotonicity Estimated costs Execution costs explain select... select... #rows,... execution time 10s 4h 10min Estimated cost diagram should always be monotonic Execution cost diagram might not be monotonic, e.g. if estimations were not precise => Even if we have Monotonic estimated cost behaviour, we want to reduce #Plans and #Segments. 16

17 Reduced Plan Diagram Plan Diagram Reduced Plan Diagram Maximum allowed Cost Increase: 5% 10% Reduced Plan Diagram: Try to reduce #Plans without increasing estimated costs more than X% in each point (X% is called Costgreedy Reduction or plan optimality tolerance threshold.) 17

18 TPC-H Query 7 Plan Diagram 4 Plans P4 covers 1,78% only Many small segments Not smooth 18 Cost Diagram monotonic Reduced Plan Diagram 2 Plans only Max Increase 6.73% Avg Increase 0.72%

19 TPC-H Query 8 Plan Diagram 9 Plans 80% covered by 3 Plans Many small segments Not smooth 19 Cost Diagram monotonic Reduced Plan Diagram 4 Plans only Max Increase 9.8% Avg Increase 0.41%

20 TPC-H Query 9 Plan Diagram 22 Plans 80% covered by 3 Plans Many small segments Not smooth Highly instable 20 Cost Diagram monotonic Reduced Plan Diagram 10 Plans Max Increase 9.15% Avg Increase 0.15%

21 Clarifications 21 #Plans does not necessarily correspond with execution time of one query. We do not rate performance! But: We consider the stability of an optimizer.

22 Hypothesis 22 Modern query optimizers make extremely fine-grained plan choices smaller plans occupy less than 1% of selectivity space Optimizers are over-sophisticated Processing overheads with query optimization could be lowered

23 Parametric query optimization (PQO) PQO at run time, use plan corresponding to selectivity parameter(s). PQO Assumptions: Plan Convexity But: today's optimizers can look like that: If Plan P is optimal at point A and at point B, it is also optimal at all points joining the two points Plan Uniqueness An optimal plan P appears at only one contiguous region in the entire space Plan Homogenity 23 apriori identify the optimal set of plans at compile time An optimal plan P is optimal within the entire region enclosed by its plan boundaries. Better: reduce #plans for PQO

24 Conclusion Often highly intricate diagrams Typically 80% of space covered by 20% of plans (postgresql: 37%) 24 Cardinality of plan diagram can be reduced, without materially affecting the query cost Assumptions of parametric query optimization literature do not hold in practice. (Reduction needed) Needed: Mechanism for pruning the plan search space Directly reduce #plans by optimizer

25 Discussion Plan Diagram Cost Diagram Do you have any questions? 25 Reduced Plan Diagram?

26 Comparison of Database Analyzers TPC-H Query postgresql OptA OptB OptC 26 %plans for 80% 100,00% 37,50% 50,00% 33,33% 13,64% 22,22% 14,29% 25,00% 37,00% Gini Index 0,47 0,65 0,57 0,68 0,76 0,48 0,24 0,28 0,52 28,7 24,5 28,8 17,00% 23,00% 16,00% 0,79 0,72 0,8 Measurement of statistical dispersion OptX: Commercial DBs tested by IIS #Plans ,13 postgresql: Same dataset, similar reference system postgresql uses smaller #Plans to cover space. postgresql needs higher percentage of plans to cover 80% of selectivity space. postgresql's Gini Index is close to 0.5 which is better than 0.7 or 0.8.

27 Generation of Reduced Plan Diagram 27 P1, P4 are upper bounds of P2 We move point to P1 which has lowest cost (90) If all Points of a plan are swit

28 Query plan costs 28 The order of the operations in a query plan is important: cost(ab) cost(abc)!= cost(a) * cost(bc) cost(abc)!= cost(ab) * cost(c)!= cost(ba)

29 Complex patterns 29 Rapidly alternating choices not created on postgresql

Analyzing Plan Diagrams of Database Query Optimizers

Analyzing Plan Diagrams of Database Query Optimizers Naveen Reddy Jayant R. Haritsa Database Systems Lab, SERC/CSA Indian Institute of Science, Bangalore 560012, INDIA Abstract A plan diagram is a pictorial